Debug a SYCL* Application on a CPU
This document describes a basic scenario of debugging a sample SYCL* app, Array Transform, with the kernel offloaded to the CPU.
Before you proceed, make sure you have completed all the necessary setup steps and successfully completed the debug session in the Get Started Guide.
Basic Debugging
NOTE:
For your convenience, all common Intel® Distribution for GDB* commands used in examples below are provided in the reference sheet.
If you have not already done so, start the debugger:
gdb-oneapi array-transform
Make sure that the kernel is offloaded to the correct device:
set env ONEAPI_DEVICE_SELECTOR=*:cpu run
Example output:
[SYCL] Using device: [Intel® Core™ i7-9750H CPU @ 2.60GHz] from [Intel® OpenCL]
Consider the Array Transform sample, which contains a simple kernel - a function that can be offloaded to different devices:
52 h.parallel_for(data_range, [=](id<1> index) {
53 size_t id0 = GetDim(index, 0);
54 int element = in[index]; // breakpoint-here
55 int result = element + 50;
56 if (id0 % 2 == 0) {
57 result = result + 50; // then-branch
58 } else {
59 result = -1; // else-branch
60 }
61 out[index] = result;
62 });
The code processes elements of the input array depending on whether they are even or odd, and produces an output array.
Define a breakpoint at line 56:
break 54
Expected output:
Breakpoint 1 at 0x405800: file /path/to/array-transform.cpp, line 54.
NOTE:
Do not expect your output to exactly match the one provided in the tutorial. The output may vary due to the nature of parallelism and different machine properties. The ellipsis […] denotes output omitted for brevity.
NOTE:
To select a CPU device we have done set env ONEAPI_DEVICE_SELECTOR=*:cpu earlier. You can check its value by issuing show env ONEAPI_DEVICE_SELECTOR.
Run the program:
run
When the thread hits the breakpoint, you should see the following output:
Starting program: <path_to_array-transform> [Thread debugging using libthread_db enabled] Using host libthread_db library "/lib/x86_64-linux-gnu/libthread_db.so.1". [New Thread 0x7ffff37dc700 (LWP 21540)] [New Thread 0x7fffdba79700 (LWP 21605)] [New Thread 0x7fffdb678700 (LWP 21606)] [New Thread 0x7fffdb277700 (LWP 21607)] [SYCL] Using device: [Intel® Core™ i7-7567U CPU @ 3.50GHz] from [Intel® OpenCL] [Switching to Thread 0x7fffdb678700 (LWP 21606)] Thread 4 "array-transform" hit Breakpoint 1, main::$_1::operator()[...] at array-transform.cpp:54 54 int element = in[index]; // breakpoint-here
Now you can issue the usual Intel Distribution for GDB commands to inspect the local variables, print a stack trace, and get information on threads. For your convenience, common Intel Distribution for GDB commands are provided in the reference sheet.
Keep debugging and display the value of the index variable:
print index
Expected output:
$1 = cl::sycl::id<1> = {24}
Continue program execution:
continue
You should see the next breakpoint hit event, which comes from another thread.
Continuing. [Switching to Thread 0x7fffdba79700 (LWP 21605)] Thread 3 "array-transform" hit Breakpoint 1, main::$_1::operator()[...] at array-transform.cpp:54 54 int element = in[index]; // breakpoint-here
If you print the value of the index variable now:
print index
The output differs from the previous one:
$2 = cl::sycl::id<1> = {32}
To print data elements, use the bracket operator of the accessor:
print in[index]
Expected output:
$3 = 132
You can also print the accessor contents in the following ways:
print in
Expected output:
$4 = {[...], MData = 0x7fffffffd3e0}x /4dw in.MData
Expected output:
0x7fffffffde30: 100 101 102 103
where the x command examines the memory contents at the given address and /4dw specifies that the examination output must contain four items in decimal format, word-length each.
Single Stepping
A common debugging activity is single-stepping in the source. The step and next commands allow you to step through source lines, stepping into or over function calls.
To check the current thread data, run the following command:
thread
You should get the following output:
[Current thread is 3 (Thread 0x7fffdba79700 (LWP 21605))]
To check the data of a particular thread, run:
info thread 3
Example output:
Id Target Id Frame * 3 Thread [...] main::$_1::operator()[...] at array-transform.cpp:54
To make Thread 3 move forward by one source line, run:
next
You should see the following output:
[Switching to Thread 0x7fffdb277700 (LWP 21607)]
Thread 5 "array-transform" hit Breakpoint 1, main::$_1::operator()[...]
at array-transform.cpp:54
54 int element = in[index]; // breakpoint-here
Stepping has not occurred. Instead, a breakpoint event from Thread 5 is received and the debugger switched the context to that thread. This happens because you are debugging a multi-threaded program and multiple events may be received from different threads. This is the default behavior, but you can configure it for more efficient debugging. To ensure the current thread executes a single line without interference, set the scheduler-locking setting to on or step. This command is useful to keep the other threads stopped while the current thread is stepping (if set to step) or resumed (if on):
set scheduler-locking step
NOTE:
The default value of scheduler-locking is replay. If you set it to on and want to resume your program with the continue command, do not forget to set scheduler-locking back to replay or off. Otherwise, only the current thread is resumed. Hence, the recommended value for scheduler-locking is step.
Continue executing the next command:
next
You should see the following output:
55 int result = element + 50;
Continue executing the next command:
next
You should see the following output:
56 if (id0 % 2 == 0) {
To see the value of index variable, run:
print index
You should see the following output:
$6 = cl::sycl::id<1> = {16}
Run:
print in[index]
The expected output looks as follows:
$7 = 116
Finally, run:
print result
You should see the following output:
$8 = 166